The contradiction is written directly into Chinese planning documents. The Fourteenth Five-Year Plan calls for constructing a green ecological civilization and achieving carbon peaking and carbon neutrality, while simultaneously orienting the entire economy around digital infrastructure, big data and AI. The plan even instructs policymakers to promote the profound convergence of the internet, big data and artificial intelligence with all industries, and to forcefully advance the coordinated transformation of industrial digitization and greening. Checketts observes that many Chinese leaders, and their counterparts in Hong Kong, do not see these aims as oppositional at all. Hong Kong’s Research Grants Committee, for instance, lists both developing a sustainable environment and big data and artificial intelligence as priority funding areas for 2026 and 2027, presenting them as twin routes to the same broad goal rather than competing demands.

China is hardly alone in this posture. The European Union wants AI leadership and climate compliance at once, Korea and Japan chase IPCC targets while courting the AI elite, and only the United States, Checketts notes, has recently pursued one goal with single-minded disregard for the other. He traces the shared confusion to the ideology of sustainable development itself, pointing out that the United Nations pairs SDG 9, which promotes industrialization and innovation, with SDG 13, which demands urgent climate action, reinforcing the belief that industrialism can be commensurate with ecological harmony. The concept of AI for sustainability adds a further layer: the claim, voiced most enthusiastically by figures such as former Google CEO Eric Schmidt, that AI is not merely compatible with fighting climate change but may be the only solution to it.

Checketts argues that this claim functions like chemotherapy rhetoric, making the planet sicker in the hope that it will heal, and he finds it potentially suicidal rather than merely optimistic. The moral hazard argument, developed in climate research by Corner and Pidgeon in their work on geoengineering, holds that when a party feels insured against an undesired outcome, it takes greater risks and may neglect urgent action. Geoengineering debates show the pattern clearly: proposals to reflect sunlight with atmospheric particulates or to sequester carbon dioxide at scale can be perceived by the public as a license to continue carbon-intensive lifestyles, producing a rebound effect that makes emissions worse. Checketts transfers the same logic to AI. AI development is currently exacerbating the climate problem, he writes, and there is no substantiated reason to believe AI will find climate solutions more efficiently than existing policy already could.

The evidence for AI’s environmental toll is substantial. Data centers already consume up to 4.4 percent of all electricity in the United States, and global data center electricity demand is projected to more than double from 415 terawatt hours in 2024 to 945 terawatt hours by 2030. Total carbon emissions from AI are estimated at around 100 megatonnes per year, a figure often compared to the global air transit industry. Training costs scale geometrically with model size: GPT-4 produced roughly 40 times the carbon of GPT-3, which itself was estimated at 550 tonnes for training, about the annual consumption of 120 American homes. Routine use is also expensive, with AI-powered search running between 5 and 25 times more inefficient than simple search algorithms, and inference demands exceeding training demands over a model’s lifetime. Company self-reporting compounds the uncertainty, since swapping energy credits has allowed tech firms to report emission rates possibly seven times lower than they actually are.

Water is the second cost. Data centers pull water from local sources to cool their heat, and while the water is not destroyed, its removal from the water table has real consequences for ecosystems and human populations. In 2023, Google alone was responsible for evaporating around 23 billion liters of water, more than PepsiCo used in the same year, and current projections estimate that by 2027 data centers will consume the equivalent of half the entire water use of the United Kingdom. Li and colleagues estimate that 10 to 50 medium-length GPT-3 responses consume around half a liter of water. In China, data centers cluster around Shanghai, Beijing, Guangzhou and Shenzhen, cities each numbering over 15 million people, and Beijing and Shanghai already face high levels of water stress. With 75 percent of Asia water insecure and countries hosting more than 90 percent of the region’s population confronting an imminent water crisis, adding thirsty data centers to the equation is a serious gamble.

The third cost is hardware, and here China’s exposure is uniquely severe. Sixty percent of all rare earth elements traded globally come from China, most from the Bayan Obo mining and refining site in Inner Mongolia. Open-pit mining there leaves elevated rare earths and heavy metals in soil up to 60 kilometers from the site, and refining requires huge amounts of water and open-air baths of sulfuric or hydrochloric acid in an arid region. Julie Klinger has described the consequences as radioactive rivers, cancer villages, acute chronic arsenic toxicity and an environmental and epidemiological crisis so grave that addressing it is now viewed as a matter of national security. Electronics manufacturing concentrated in southern China adds toxic fumes, soil pollution and groundwater poisoning, with fine particulate air pollution contributing to one-fifth of all premature deaths worldwide and rates as high as 26 percent in South and East China. Components in data centers last only 3 to 5 years, yet only around 20 percent of electronics are properly recycled, leaving an estimated 63 million tonnes of e-waste globally in 2022, much of it historically routed through southern China and now increasingly to South Asia.

What makes China’s situation distinctive is that its own doctrine appears to condemn precisely this pattern. The Ten Definites articulated by Xi Jinping state that the principal contradiction in China’s society in the New Era is between the people’s growing need for a better life and unbalanced and inadequate development, demanding a people-centered development ideology. The Scientific Outlook on Development warns against economic growth relying mainly on expanded investment scale, resulting in excessive resource and environmental costs, and calls for a path characterized by high technological content, low resource consumption and minimal environmental pollution. The Fourteen Commitments include improving people’s well-being as the primary goal of development, and the ninth commitment insists that lucid waters and lush mountains are invaluable assets and that the ecological environment must be treated as life itself. Checketts argues that AI growth pursued at all costs, with proliferating data centers, stressed water supplies and expanding pollution, constitutes unbalanced development that these commitments were designed to prevent.

His prescription is not to abandon AI but to discipline it. China should prioritize hardware efficiency and slimmer AI models over data center proliferation, direct development toward specific applications in agriculture, energy efficiency and healthcare rather than a blanket promotion of the technology, and avoid the moral hazard of trusting AI for sustainability as insurance against the climate crisis. The report that DeepSeek required far fewer resources than comparable large language models is significant on this front, though Checketts notes that benchmarking work suggests DeepSeek may actually consume the most energy in real-world performance, and that overuse of LLMs for mundane purposes constitutes significant waste regardless. He points to signs of a different path: China has promised to expand renewable energy sources sixfold by 2035, aims to peak carbon emissions before 2030, and appears more focused on practical AI applied to national solutions than on the promethean artificial general intelligence project pursued by American companies.

The deeper question the paper raises is whether climate change is fundamentally a technological problem at all. Checketts, drawing on Herbert Marcuse, suggests that in economies prioritizing consumption over balance, technological fixes may only pave the way for new demands, and that what is truly needed is a different orientation toward material production and consumption. As a communist power whose founding tenets prioritize economic development for the people rather than for the sake of growth, China has, in his view, a real opportunity to demonstrate that AI can be developed within specific parameters and limitations rather than for unchecked profit. Whether it seizes that opportunity remains uncertain, but the stakes, for China and for the world watching its model, could hardly be higher.

Subject of Research: The environmental costs of AI development in China and the contradiction with Chinese sustainability policy

Article Title: The sustainable AI development contradiction: Chinese policy on sustainability and the costs of AI development

Article References: Checketts, L. (2026). The sustainable AI development contradiction: Chinese policy on sustainability and the costs of AI development. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03296-z

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03296-z

Keywords: artificial intelligence, China, climate change, sustainability, data centers, moral hazard, carbon emissions, water consumption, rare earth elements, e-waste, Chinese policy, AI & Society